Floor 8: Production (AgentCore)

Files on this floor: 8_agentcore_production/.env.example

CORDULA’S LOG, FLOOR 8. The last floor is the corporate one: glass towers across Lake Washington, lanyards, a Regional Managing Director who wants us gone by end of quarter, and, past all of it, the Tacoma Narrows, where the old giant octopus legend lives. Everything we built so far stops the moment someone closes a laptop. Down here it has to keep running on its own. So we moved it, one piece at a time, into the cloud: the dice, the registration desk, the rule book, the Narrator itself. Barnaby insisted on doing it in order. He has always respected a checklist.

System AI: “Floor 8. Production. Real AWS resources, real IAM roles, real invoices. This is the part of the crawl where mistakes cost money. Set your region before you do anything else. I mean it.”

Floor Objective

Take your game to production, one brick at a time: the dice server, the Character Agent, the Rules Agent and finally the Narrator move to Amazon Bedrock AgentCore. Each section of this floor moves one brick to the cloud while the others keep running on your machine, so the game keeps working at every step.

What is AgentCore Runtime? AgentCore Runtime is a managed AWS service that hosts agents and MCP servers. You hand it your code; it packages it, runs each session in its own isolated microVM, scales with the load, authenticates every caller with IAM and sends the logs to Amazon CloudWatch. You manage no server, no container and no load balancer. The AgentCore CLI (agentcore) drives the whole process from your terminal: create a project, declare what to deploy, deploy, test, read the logs.

The plan. Everything lives in a single AgentCore project named gameprod, created on this page. Each section adds one brick to it, and each brick has its own folder in 8_agentcore_production:

The Narrator (narrator/narrator_orchestrator.py) is the one file you come back to in every section: each time a brick moves to the cloud, the Narrator is pointed at it. Its TODOs therefore carry the number of their section (8.1 Step 1, 8.2 Step 1…). The code is the code you already wrote, with a few lines changed. This floor builds on Floors 4 to 6 only: Floor 7 (Strands harness) is not needed.

The code to write is given. Under each TODO, the sections show the code to paste, with the TODO line above it so you can see where it goes; keep the TODO in your file. When a step replaces or deletes code, a Before / After pair shows the change. The explanation under each block says why the code looks the way it does.

SYSTEM WARNING: Where to run this floor: it deploys real resources to your own AWS account: IAM roles, CloudFormation stacks and AgentCore resources. Use credentials allowed to create IAM roles, CloudFormation stacks and AgentCore resources. The commands use us-west-2; any AgentCore region works: set yours in the export of Step 1.

What Floor 8 Costs

Floors 0 to 7 bill only for model calls. Floor 8 also leaves resources in your account until the Cleanup page removes them. Most bill only when they are used; a few bill a little for as long as they exist. The rates below are the us-west-2 list prices as of November 2026, from the AgentCore pricing page, the Amazon Bedrock pricing page (Managed Knowledge Bases) and the CloudWatch pricing page. The AgentCore Runtime rates are also quoted in the AWS sample Cost and Reliability guide. Check them for your region before you start.

The worked numbers: 0.5 GB x 0.25 hour x $0.00945 = $0.0012 for an idle session, and 13 KB x $5.00 per GB = $0.00006 a month for the handbook. The token counts and event counts are estimates for one learner playing a few dozen turns. A whole run of Floor 8 should stay under five dollars, most of it model calls. Your real numbers show up in AWS Cost Explorer, grouped by service, about a day later.

Nothing here has a monthly minimum, and nothing bills by the hour while idle except storage. That is not a reason to keep it: the Cleanup page removes every row of this table.

Set Up the Project

System AI: “Everything from here costs money in somebody’s AWS account. Possibly yours. Set AWS_DEFAULT_REGION in every terminal. The System will not remind you twice. The System has now reminded you twice.”

Make sure you are at the root of Floor 8 in your terminal:

cd aiRPG-seattle-dungeon-crawl/8_agentcore_production

Step 1: Install the AgentCore CLI

ITEM ACQUIRED: [agentcore CLI 0.31.1]

Rarity: Pinned. Version 1.0 renames several commands. Stay on 0.31.1 for the whole floor or explain yourself to the audience.

The CLI needs Node.js 20 or later. Install the version this crawl was written against:

npm install -g @aws/agentcore@0.31.1
agentcore --version

agentcore --version should print 0.31.1. The CLI evolves quickly and version 1.0 renames several commands, so stay on this version for the whole floor. It uses your usual AWS credentials, like the agents of the previous floors. Set the region once, in every terminal you use for this floor:

export AWS_DEFAULT_REGION=us-west-2

The CLI deploys to this region, and the code of the floor never names one: boto3, the AWS SDK for Python, reads AWS_DEFAULT_REGION, and every client of the floor takes its region from there. The variable takes precedence over the region of your AWS profile; without it, the CLI deploys to the region of your profile, and the Narrator looks for your runtimes there. Once deployed, the agents get the region of their runtime from AgentCore.

Check: agentcore --version prints 0.31.1, and aws sts get-caller-identity prints your account id. If agentcore is not found, open a new terminal: npm puts global commands on your PATH, and an older terminal does not see them. If npm install -g fails with EACCES, your npm installs global packages in a system folder: use a Node.js version manager such as nvm, or set an npm prefix in your home folder.

CORDULA: Barnaby made me read the version number out loud, twice, and write it on the back of my hand. “We will not upgrade mid-operation, Sergeant.” I did not argue. I have watched a firmware update brick a landing craft on a beach.

Step 2: Create the Project

Run this command from 8_agentcore_production, the folder that holds diceserver/ and narrator/:

agentcore create --project-name gameprod --no-agent

--no-agent creates an empty project: we will add our own code instead of a generated template. The new gameprod/ folder sits next to the brick folders. Have a look at gameprod/agentcore/agentcore.json: this file describes everything the project deploys, and it is empty for now. The gameprod/agentcore/cdk/ folder holds the infrastructure code the CLI generates for you; you never edit it.

Step 3: Start the Knowledge Base

SPONSOR MESSAGE: Bedrock indexing is brought to you by patience: a few minutes in your own account, longer in a sandbox with a low embedding quota. Start it now and go deploy some dice.

In 8.3, the Rules Agent searches the Crawl Handbook in an Amazon Bedrock knowledge base. Bedrock needs time to index it: the handbook is eleven small Markdown files, so a few minutes, longer in an account with a low embedding quota. Start it now: the indexing runs in the background while you work through 8.1 and 8.2. Here you only run the commands; 8.3 explains them.

The handbook is the one the Floor 5 knowledge base was built from, in workshop/world/crawl-handbook/. Run the commands from the project folder:

cd gameprod
BUCKET=gameprod-rules-$(aws sts get-caller-identity --query Account --output text)
aws s3 mb s3://$BUCKET
aws s3 cp ../../workshop/world/crawl-handbook/ s3://$BUCKET/ --recursive
agentcore add knowledge-base --name rules --description "Crawl Handbook" --source s3://$BUCKET/
agentcore deploy -y

aws s3 cp --recursive should list eleven uploads, one per handbook section. The files go into a new S3 bucket, the project declares a knowledge base that reads it, and agentcore deploy creates it in a few minutes. When the deployment ends, the knowledge base shows Deployed but is not ready yet: Bedrock goes on indexing the handbook. Nothing to wait for: 8.3 checks the indexing.

These commands create resources in your AWS account: an S3 bucket and a knowledge base. Remove them at the end of the workshop, as described on the Cleanup page.

Check: agentcore status lists rules under Knowledge Bases as Deployed, with Ingestion: IN_PROGRESS or COMPLETE. If aws s3 mb fails with BucketAlreadyExists, the name is taken: bucket names are global, so add a suffix to BUCKET and keep it for the Cleanup page. If agentcore deploy fails, read 8.1, Part 2, Step 2: its warnings cover the usual errors.

CORDULA: We dropped the rule book into a bucket and walked away from it, which felt wrong. Bellevue is the kind of floor where the paperwork takes longer than the work. Barnaby called it a “long-lead item” and moved on to the dice.

The project is ready: head to 8.1 The Dice Server to deploy its first brick.

Boss Battle: The Quarterly Earnings Call (optional)

CORDULA: Bellevue has conference rooms the way the estuary has tide pools. In one of them, the accountant from Floor 5 had dialed in a banker and the Regional Managing Director, and we were the only item on the agenda.

The Quarterly Earnings Call: the accountant reviews, the banker spins, the Regional Managing Director signs off. A Strands graph with fixed edges and a condition on each one: the call only goes on if the party took the last round. It starts from the Floor 5 code and runs on your machine, so you can play it before, during or after the deployments of this floor.

SYSTEM DESIGN, LAYER 8: A PERSISTENT WORLD. A LitRPG world exists whether or not the reader is looking. On your laptop, the System ends when you close the terminal. On AgentCore, the dice, the registry, the rules and the Narrator each run on their own, the characters live in AgentCore Memory, the rules in a managed knowledge base, and every turn leaves a trace. That is the difference between a game you run and a world that runs: Cordula’s sheet is still there tomorrow, and so is the accountant’s grudge. See The System.

NEW ACHIEVEMENT UNLOCKED!

[WE DEPLOYED ON A FRIDAY]

Reward: You received a [Gold Box]. Inside: an empty AgentCore project, a knowledge base quietly indexing in the background, and a sense of foreboding.

System AI: “The project exists. Deploy the first brick, then the next. I will be watching your CloudFormation events.”